Learning Deep Structured Models

Authors: Liang-Chieh Chen, Alexander Schwing, Alan Yuille, Raquel Urtasun

ICML 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the effectiveness of our algorithm in the tasks of predicting words from noisy images, as well as tagging of Flickr photographs. We show that joint learning of the deep features and the MRF parameters results in significant performance gains.
Researcher Affiliation Academia University of California Los Angeles, USA; University of Toronto, 10 King s College Rd., Toronto, Canada
Pseudocode Yes Figure 1: Algorithm: Deep Structured Learning, Figure 3: Algorithm: Efficient Deep Structured Learning
Open Source Code Yes The library is released on http://alexander-schwing.de.
Open Datasets Yes We took the lower case characters from the Chars74K dataset (de Campos et al., 2009)... We initialized the deep-net parameters using a model pre-trained on Image Net (Deng et al., 2009).
Dataset Splits Yes The training, validation and test sets have 10, 000, 2, 000 and 2, 000 variations of words respectively. For all experiments, the validation set is only used to decrease the step size, i.e., if the accuracy on the validation set decreases, we reduce the step size by 0.5.
Hardware Specification No The paper states, 'It supports usage of the GPU for the forward and backward pass,' but does not provide specific hardware models (e.g., GPU model, CPU type, or memory size).
Software Dependencies No The paper mentions C++, HDF5 storage, and Google protocol buffers, but does not provide specific version numbers for any of these software dependencies.
Experiment Setup Yes In particular, we use a mini-batch size of 100, a step size of 0.01 and a momentum of 0.95. If the unary potential is pre-trained, the initial step size is reduced to 0.001. All the unary classifiers are trained with 100, 000 iterations over mini-batches. We use ϵ = 1, set cr = 1 for all regions r, and perform 10 message passing iterations.